Main effects and interactions

Hello All,
I have a problem on my hands that I was wondering your thoughts on. My
team has run a series of logistic regressions predicting a host of
psychological disorders (outcome is dichotomous: yes/no have disorder)
from a host of independent predictors. Initially we thought that we
would find an interaction between two of our main predictors (site X
distress).

That was not the case, instead we just found Distress to be significant
That is distress predicted a host of psychological disorders above and
beyond everything else, and site (or rather place of birth) did not
predict anything. We decided to keep site in the model as this was of
central theoritical importance but we dropped the interaction between
site and distress.

HOWEVER, when we ran the reduced model, and dropped the interaction, we
now find that the main effect for distress was not signiciant for one
of the disorders. In addition, its effect is reduced for the other
disorders, although it still remains significant.

Why would this be? I can't understand how dropping an insignificant
interaction would change the significance of a previously robust main
effect. I am also wondering which model I should present - the full
model with the interaction or the reduced model.

Thanks
Irene

.